update

basedformer/lm_utils.py

@@ -28,7 +28,7 @@ def init_weights(model, n_layer):
 
 def init(model_class, config):
     model = model_class(config)
-    model.init_weights()
+    init_weights(model, config["n_layer"])
     return model
 
 def no_init(model_class, config):

basedformer/models/__init__.py

@@ -2,12 +2,15 @@ from . import gptj
 from . import gpt2
 from . import fairseq
 from . import gptneo
+from . import alibi
+from . import fast
 
 MODEL_MAP = {
     "gptj": gptj.GPTJModel,
     "gpt2": gpt2.GPT2Model,
     "gpt-fairseq": fairseq.GPTFairModel,
-    "gpt-neo": gptneo.GPTNeoModel
+    "gpt-neo": gptneo.GPTNeoModel,
+    "alibi": alibi.AlibiModel,
 }
 
 def get_model(model_name: str):

basedformer/models/alibi.py

+from typing import Callable, KeysView
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from basedformer.utils import *
+from torch.utils.checkpoint import checkpoint as ck
+from einops import rearrange, repeat
+try:
+    from collections.abc import MutableMapping
+except ImportError:
+    from collections import MutableMapping
+import os
+from pathlib import Path
+import math
+from basedformer.models import base_lm
+
+def get_slopes(n):
+    def get_slopes_power_of_2(n):
+        start = (2**(-2**-(math.log2(n)-3)))
+        ratio = start
+        return [start*ratio**i for i in range(n)]
+
+    if math.log2(n).is_integer():
+        return get_slopes_power_of_2(n)                   #In the paper, we only train models that have 2^a heads for some a. This function has
+    else:                                                 #some good properties that only occur when the input is a power of 2. To maintain that even
+        closest_power_of_2 = 2**math.floor(math.log2(n))  #when the number of heads is not a power of 2, we use this workaround. 
+        return get_slopes_power_of_2(closest_power_of_2) + get_slopes(2*closest_power_of_2)[0::2][:n-closest_power_of_2]
+
+def _attn(query, key, value, causal_mask, masked_bias,
+            attention_mask=None, scale_attn=None):
+
+    attn_weights = torch.matmul(query, key.transpose(-1, -2))
+    attn_weights = torch.where(causal_mask, attn_weights, masked_bias.to(attn_weights.dtype))
+    attn_weights = attn_weights / scale_attn
+
+    if attention_mask is not None:
+        attn_weights = attn_weights + attention_mask
+
+    attn_weights = F.softmax(attn_weights, dim=-1)
+    attn_weights = attn_weights.to(value.dtype)
+
+    attn_output = torch.matmul(attn_weights, value).to(value.dtype)
+
+    return attn_output
+
+class SelfAttention(nn.Module):
+    # Code copied from HF, might want to sanity check later.
+    def __init__(self, config):
+        nn.Module.__init__(self)
+        max_positions = 2049
+        bias = torch.tril(torch.ones((max_positions, max_positions), dtype=torch.uint8, requires_grad=False)).view(
+            1, 1, max_positions, max_positions).bool()
+
+        self.register_buffer("slopes", torch.Tensor(get_slopes(config.n_head)))
+        #In the next line, the part after the * is what constructs the diagonal matrix (right matrix in Figure 3 in the paper). 
+        #If you run it you'll see that it doesn't exactly print out the same matrix as we have in Figure 3, but one where all rows are identical.
+        #This works because the softmax operation is invariant to translation, and our bias functions are always linear.
+        print(self.slopes.shape)
+        self.alibi = self.slopes.unsqueeze(1).unsqueeze(1) * torch.arange(max_positions).unsqueeze(0).unsqueeze(0).expand(config.n_head, -1, -1)
+        self.alibi = self.alibi.view(config.n_head, 1, max_positions)
+
+        self.head_dim = config.hidden_dim // config.n_head
+        self.rotary_dim = self.head_dim // 4
+        self.hidden_dim = config.hidden_dim
+        self.n_head = config.n_head
+        self.q_only = config.q_only
+        self.register_buffer("scale_attn", torch.sqrt(torch.tensor(self.head_dim, requires_grad=False).float()))
+        self.register_buffer("bias", bias)
+        self.register_buffer("masked_bias", torch.tensor(-1e9, requires_grad=False)) #-1e10 is what mtj uses.
+        attn_bias = False
+        if config.q_only:
+            self.k_proj = nn.Linear(self.hidden_dim, self.head_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
+            self.v_proj = nn.Linear(self.hidden_dim, self.head_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
+        else:
+            self.k_proj = nn.Linear(self.hidden_dim, self.hidden_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
+            self.v_proj = nn.Linear(self.hidden_dim, self.hidden_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
+
+        self.q_proj = nn.Linear(self.hidden_dim, self.hidden_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
+        self.out_proj = nn.Linear(self.hidden_dim, self.hidden_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
+
+    def forward(self, x, kv=None, cache=False):
+        B, S, H = x.shape # batch, sequence, hidden_dim
+        # split heads into: [batch, head, sequence, head_dim]
+        # transpose q, k after rotary as rotary code accepts [b, s, h, h_d]
+        query = self.q_proj(x).view(B, S, self.n_head, self.head_dim).transpose(1, 2)
+        if self.q_only:
+            key = self.k_proj(x).view(B, S, 1, self.head_dim).transpose(1, 2)
+            value = self.v_proj(x).view(B, S, 1, self.head_dim).transpose(1, 2) 
+        else:
+            key = self.k_proj(x).view(B, S, self.n_head, self.head_dim).transpose(1, 2)
+            value = self.v_proj(x).view(B, S, self.n_head, self.head_dim).transpose(1, 2)
+
+        if kv:
+            offset = kv[0].shape[-2]
+        else:
+            offset = 0
+
+        self.alibi = self.alibi.repeat(B, 1, 1)  # batch_size, 1, 1
+        
+        if kv:
+            k, v = kv
+            # cat key and value (get the whole sequence, other than the last added token all are cached),
+            # so query can attend to it.
+            key = torch.cat([k, key], dim=-2) # cat key
+            value = torch.cat([v, value], dim=-2) # cat value
+
+        query_length, key_length = query.size(-2), key.size(-2)
+        #causal mask with generation in mind
+        causal_mask = self.bias[:, :, key_length - query_length:key_length, :key_length]
+        print(causal_mask.shape)
+        print(self.alibi.shape)
+
+        x = _attn(
+            query, key, value, causal_mask+self.alibi, self.masked_bias, None, self.scale_attn
+        )
+
+        x = x.transpose(1, 2).contiguous().view(B, S, H)
+        x = self.out_proj(x)
+
+        if cache:
+            return x, (key, value)
+        else:
+            return x, None
+
+class FeedForward(nn.Module):
+    def __init__(self, config):
+        nn.Module.__init__(self)
+        self.ff1 = nn.Linear(config.hidden_dim, config.hidden_dim*4, device=config.device, dtype=config.dtype)
+        self.ff2 = nn.Linear(config.hidden_dim*4, config.hidden_dim, device=config.device, dtype=config.dtype)
+        self.activation = config.activation
+
+    def forward(self, x, act_ck=False):
+        x = self.ff1(x)
+        if act_ck:
+            x = ck(self.activation, x)
+        else:
+            x = self.activation(x)
+        x = self.ff2(x)
+        return x
+
+class AlibiLayer(nn.Module):
+    def __init__(self, attn, ff, config):
+        nn.Module.__init__(self)
+        self.ln_preattn = nn.LayerNorm(config.hidden_dim, eps=config.eps, device=config.device, dtype=config.dtype)
+        #self.ln_preattn = nn.LogSoftmax(dim=-2)
+        self.ff = ff(config)
+        self.attn = attn(config)
+        self.tick = True
+
+    def forward(self, x, layer_id=None, hypernetwork=None, act_ck=False, diff_hypernets=False, interleaving_layers=False, every_n=5, cache=False, kv=None):
+        residual = x
+        if act_ck:
+            x = ck(self.ln_preattn, x)
+            attn_out, kv = ck(self.attn, x, kv, cache)
+            #attn_out, kv = self.attn(x, kv=kv, cache=cache)
+
+        else:
+            x = self.ln_preattn(x)
+            attn_out, kv = self.attn(x, kv=kv, cache=cache)
+
+        if hypernetwork:
+            if diff_hypernets:
+                if interleaving_layers and layer_id % every_n == 0:
+                    if self.tick:
+                        hyper_out = hypernetwork[0](x)
+                        self.tick = False
+                    else:
+                        hyper_out = hypernetwork[1](x)
+                        self.tick = True
+
+                elif layer_id % every_n == 0:
+                    hyper_out = hypernetwork[(layer_id // every_n) - 1](x)
+
+            else:
+                if layer_id % every_n == 0:
+                    hyper_out = hypernetwork(x)
+
+        ff_out = self.ff(x, act_ck)
+        #order of addition matters, i had no idea... fixed a bug here.
+        x = attn_out + ff_out + residual
+        #x = residual + attn_out + ff_out -> doesn't match.
+        if hypernetwork and layer_id % every_n == 0:
+            x = x + hyper_out
+            
+        return x, kv
+
+class AlibiModel(base_lm.BaseModel):
+    def __init__(self, user_config, **kwargs):
+        self.default_config = {
+            'n_layer': 6,
+            'n_head': 8,
+            'n_tokens': 2048,
+            'hidden_dim': 512,
+            'vocab_dim': 50400,
+            'eps': 1e-5,
+            'device': torch.device('cuda'),
+            'dtype': torch.float16,
+            'Layer': AlibiLayer,
+            'activation': gelu_new,
+            'SelfAttention': SelfAttention,
+            'FeedForward': FeedForward,
+        }
+        base_lm.BaseModel.__init__(self, user_config, **kwargs)

basedformer/models/gptj.py

@@ -59,12 +59,18 @@ class SelfAttention(nn.Module):
         self.rotary_dim = self.head_dim // 4
         self.hidden_dim = config.hidden_dim
         self.n_head = config.n_head
+        self.q_only = config.q_only
         self.register_buffer("scale_attn", torch.sqrt(torch.tensor(self.head_dim, requires_grad=False).float()))
         self.register_buffer("bias", bias)
         self.register_buffer("masked_bias", torch.tensor(-1e9, requires_grad=False)) #-1e10 is what mtj uses.
         attn_bias = False
-        self.k_proj = nn.Linear(self.hidden_dim, self.hidden_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
-        self.v_proj = nn.Linear(self.hidden_dim, self.hidden_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
+        if config.q_only:
+            self.k_proj = nn.Linear(self.hidden_dim, self.head_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
+            self.v_proj = nn.Linear(self.hidden_dim, self.head_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
+        else:
+            self.k_proj = nn.Linear(self.hidden_dim, self.hidden_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
+            self.v_proj = nn.Linear(self.hidden_dim, self.hidden_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
+
         self.q_proj = nn.Linear(self.hidden_dim, self.hidden_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
         self.out_proj = nn.Linear(self.hidden_dim, self.hidden_dim, bias=attn_bias, device=config.device, dtype=config.dtype)
         sin, cos = fixed_pos_embedding(dim=self.rotary_dim, seq_len=max_positions)
@@ -76,8 +82,12 @@ class SelfAttention(nn.Module):
         # split heads into: [batch, head, sequence, head_dim]
         # transpose q, k after rotary as rotary code accepts [b, s, h, h_d]
         query = self.q_proj(x).view(B, S, self.n_head, self.head_dim)
-        key = self.k_proj(x).view(B, S, self.n_head, self.head_dim)
-        value = self.v_proj(x).view(B, S, self.n_head, self.head_dim).transpose(1, 2)
+        if self.q_only:
+            key = self.k_proj(x).view(B, S, 1, self.head_dim)
+            value = self.v_proj(x).view(B, S, 1, self.head_dim).transpose(1, 2) 
+        else:
+            key = self.k_proj(x).view(B, S, self.n_head, self.head_dim)
+            value = self.v_proj(x).view(B, S, self.n_head, self.head_dim).transpose(1, 2)
 
         if kv:
             offset = kv[0].shape[-2]
@@ -147,6 +157,7 @@ class GPTJLayer(nn.Module):
     def __init__(self, attn, ff, config):
         nn.Module.__init__(self)
         self.ln_preattn = nn.LayerNorm(config.hidden_dim, eps=config.eps, device=config.device, dtype=config.dtype)
+        #self.ln_preattn = nn.LogSoftmax(dim=-2)
         self.ff = ff(config)
         self.attn = attn(config)
         self.tick = True

scripts/cudagraph.py

@@ -5,7 +5,7 @@ from time import perf_counter, perf_counter_ns
 import numpy as np
 from tqdm import tqdm
 from contextlib import contextmanager
-from basedformer.hypernet import * 
+from basedformer.models.hypernet import * 
 import sys
 #replicating timeit magic function of ipython
 def timeit(func, r=1, n=5, quiet=False, function=None, do_tqdm=False, first=True):

scripts/q_only.py

+from basedformer import models, utils
+import torch
+config = {
+    "n_layer": 28,
+    "n_head": 16,
+    "hidden_dim": 4096,
+}
+
+config = {
+    "n_layer": 40,
+    "n_head": 40,
+    "hidden_dim": 5120,
+}
+
+config_q = {**config, "q_only":True}
+#init param matched GPT
+gpt = models.fairseq.GPTFairModel(config).cuda().half()
+utils.print_parameters(gpt)
+bsz = 3
+cached_seq = 1000
+y = torch.randint(0, 50256, (bsz, cached_seq)).long().cuda()
+x = torch.randint(0, 50256, (bsz, 1)).long().cuda()
+cache_f = torch.rand(bsz, config["n_head"], cached_seq, config["hidden_dim"]//config["n_head"]).cuda().half()
+cache_f = (cache_f, cache_f)
+cache_f = [cache_f for _ in range(config["n_layer"])]
+print(len(cache_f))
+print(cache_f[0][1].shape)
+######
+cache_q = torch.rand(bsz, 1, cached_seq, config["hidden_dim"]//config["n_head"]).cuda().half()
+cache_q = (cache_q, cache_q)
+cache_q = [cache_q for _ in range(config["n_layer"])]
+print(cache_q[0][0].shape)
+with torch.no_grad():
+    #print("Initial Context GPT:")
+    #utils.timeit(func=lambda: gpt(y), r=10, n=10)
+    out = gpt(y, cache=True)
+    print(out[1][0][0].shape)
+    print("GPT")
+    utils.timeit(func=lambda: gpt(x, kv=cache_f), r=10, n=10)
+
+
+'''
+del gpt
+#init param matched Q-Only
+gpt_q = models.gptj.GPTJModel(config_q).cuda().half()
+utils.print_parameters(gpt_q)
+
+with torch.no_grad():
+    #print("Initial Context GPT-Q:")
+    #utils.timeit(func=lambda: gpt_q(y), r=10, n=10)
+    out_q = gpt_q(y, cache=True)
+    print("GPT-Q:")
+    utils.timeit(func=lambda: gpt_q(x, kv=cache_q), r=10, n=10)
+'''
\ No newline at end of file

train.py

@@ -5,7 +5,7 @@ import torch.cuda.amp as amp
 import torch.optim as optim
 from pathlib import Path
 from torch.utils import data
-from basedformer import optimizer, utils, gptj, noemblm, gpt2
+from basedformer import optimizer, utils, models, lm_utils
 import yaml
 import sys
 from tqdm import tqdm
@@ -14,12 +14,17 @@ import wandb
 import numpy as np
 import os
 
+def softmax_activation(x):
+    return F.log_softmax(x, dim=-1)
+
 model_config = {
-    "n_layer": 3,
-    "n_head": 16,
-    "hidden_dim": 1024,
+    "n_layer": 12,
+    "n_head": 12,
+    "hidden_dim": 768,
     "vocab_dim": 50400,
     "eps": 1e-5,
+    "q_only": True,
+    "activation": torch.nn.GELU(),
 }
 
 # we need 250 batch size to train the small GPT.
@@ -29,9 +34,9 @@ train_config = {
     #"data_path": "/home/xuser/diffusionstorage/datasets/sigurd/map/sigurd_v5_fs_2049.map",
     "save_path": "/home/xuser/diffusionstorage/workspace/kuru/basedformer/models/owt2gptj-nopos-tokenshift-superhighlr-residualgate-3L-16A-1024H",
     "do_save": False,
-    "run_name": "gptj-nopos-tokenshift-superhighlr-owt2-72M-3L-16A-1024H-fp16AMP-512ctx-16bs-1e-4lrinit",
-    "lr": 1e-3,
-    "end_lr": 1e-3,
+    "run_name": "gptj-owt2-512ctx-12L-12H-768H-16bs-1e-4lr-q-only-smallattneveryotherlayer",
+    "lr": 1e-4,
+    "end_lr": 1e-4,
     "warmup_steps": 100,
     "bs": 16,
     "gas": 1,
@@ -47,7 +52,8 @@ gas = train_config["gas"]
 Path(train_config["save_path"]).mkdir(parents=True, exist_ok=True)
 
 #model = GPTModel.gpt2_init(model_config).cuda().float()
-model = noemblm.GPTNEBaseLM.init(model_config).cuda().float()
+model = lm_utils.init(models.fast.GPTJModel, model_config).cuda().float()
+utils.print_parameters(model)
 model.train()
 
 cp_list = sorted(os.listdir(train_config["save_path"]), key=lambda x: int(x.split("_")[-1]))
@@ -87,7 +93,7 @@ for input_ids, labels in t:
     loss = 0
     for x in range(train_config["gas"]):
         with torch.cuda.amp.autocast(enabled=train_config["amp"], dtype=torch.float16):
-            logits = model.lm(input_ids[x*bs:(x+1)*bs, :512].cuda(), act_ck=False)
+            logits = model(input_ids[x*bs:(x+1)*bs, :512].cuda(), act_ck=False)
             #print(tokenizer.decode(input_ids[x*bs:(x+1)*bs, :][0]))
             #roll down the sequence
             logits = logits.view(-1, logits.shape[-1])
@@ -117,7 +123,7 @@ for input_ids, labels in t:
     opt.zero_grad()
     sec_per_step = (time.perf_counter() - timex)
     step_per_sec = (1. / sec_per_step)
-    tokens_per_sec = (step_per_sec * 1024) * bs * gas
+    tokens_per_sec = (step_per_sec * 512) * bs * gas
     t.set_description(f"{step_per_sec:.2f} steps/s, {sec_per_step:.2f}s/step, {tokens_per_sec:.2f}tokens/s, loss={loss:.4f}")
     wandb.log(
         {